Research On Theory And Application Of Shape Representation And Matching

Shape representation and shape matching are the fundamental issues in the field of computer vision. Shape of the object can convey large amount of information. People tend to recognize objects according to their shapes.The traditional view is that contour provides enough information to describe the shape characteristics of the objects. For simple object, its contour reflects sufficiently the overall information of the object; for the complex ones, it is difficult to adequately describe the object using its contour alone. As is well-known, the feature of object is described not only by the boundary, region infounation surrounded by the contour is also very important. Consequently, combing the contour and region information can make a better description of shape.This work is divided into two parts. Based on the analysis of existing methods, we do some experiments on two types of images:remote sensing images (rich in texture and geometric information) and simple skeleton images (rich in geometric information). Main results are listed as follows: 1. Building recognition in remote sensing imageThis paper presents a method for buildings recognition based on the rectangular prior of the building. Firstly, we detect lines using the LSD(Line Segment Detector)algorithm, and then construct rectangles according to the specified constraints and the spatial relationship among lines; then the features of color, texture and boundary are combined together to train the classifier using the GentleBoost algorithm; in the end, we deal the occluded rectangles to obtain the recognition result. 2. Shape based recognition on skeleton imageBased on the analysis of the existing shape descriptors, we do some research on shape context, a novel shape descriptor in which shapes are represented by a set of points sampled from the shape contours. We demonstrate shape recognition in the MPEG-7 shape database and some CAD drawings making use of the shape context. Furthermore, we compare the results with the HoG descriptor.